Nvidia Unveils Hardware-Isolated Security Controls for Autonomous AI Agents
Emerging Technology Risk · 9 sources · Updated 24d ago · First seen Oct 21, 2026

Nvidia presented a hardware-backed reference architecture for its Open Agent Safety Platform at GTC in Berlin, extending the open-source OpenShell runtime with independent monitoring for autonomous AI agents. The design separates model inference on Vera Rubin GPUs, agent execution on Vera CPUs, and Sentry security monitoring on BlueField-4 DPUs. Sentry is intended to enforce access policies independently of both the language model and agent sandbox; Nvidia claims it can isolate and stop agents within milliseconds when they cross configured security boundaries.
OpenShell controls access to files, programs, network services, and inference endpoints. Its policy pipeline compiles organizational policies into OPA/Rego enforcement rules and SMT-LIB representations, which the Z3 theorem prover checks for permission-ceiling violations. These checks do not establish least privilege or semantic safety: agents can still cause harm through authorized access, and verification does not cover implementation flaws or unmodeled behavior. OpenShell v0.1.0 is available under Apache 2.0 and runs on any hardware, while Sentry remains a reference design with general availability planned for early 2027. Organizations evaluating the platform should distinguish verified permission boundaries from assurance that an agent’s permitted actions are safe.


